Ask HN: Lisp Macros for Self-Improving AI? | Hacker News Reader

Ask HN: Lisp Macros for Self-Improving AI?

17 pointsbehnamoh10 comments

I'm not an expert in Lisp but from what I understand, Lisp macros allow for self-modifying code while the program is running:

https://stackoverflow.com/questions/36416532/lisp-in-ai-self-modifying-code-and-domain-specific-levels-of-abstraction

I was wondering if this is useful for AI that can modify (improve) its own code. Am I missing something?

View on HN

Self-modifying code (whether it's in Lisp, Java, or "other") could theoretically be useful, but there's a lot of "the devil's in the details" stuff to consider.

The first thing that comes to mind is the general idea that for any kind of continually improving system (the obvious analogy would be an evolutionary algorithm[0] of some sort) you need a "fitness function" that dictates whether or not the system is, in fact, improving. And defining the fitness function and the representation that maps from your code to the thing you're trying to improve, is often the hardest part.

FWIW, people have explored the intersection of evolutionary algorithms and code generation under the rubric of "Genetic Programming"[1][2][3]. This isn't necessarily the only (or best) such approach, but it's one that has been explored quite extensively and where there is a large body of literature to review. You might find some of that stuff interesting, vis-a-vis this overall topic.

[0]: https://en.wikipedia.org/wiki/Evolutionary_algorithm

[1]: https://en.wikipedia.org/wiki/Genetic_programming

[2]: https://geneticprogramming.com/

[3]: http://www.genetic-programming.org/

Other languages are similarly dynamic. For instance a Java program can write Java bytecode and pass it to the the defineClass() method of the classloader and then run the code.

Currently fashionable machine learning approaches don't generate general purpose code but instead determine the coefficients for a matrix equation. Notably this kind of "program" is guaranteed to finish, which limits what it can do algorithmically, see

https://en.wikipedia.org/wiki/BlooP_and_FlooP

but that kind of program can run efficiently on a GPU or similar device because it is a bunch of repetitive math.

There are some older algorithms like

https://en.wikipedia.org/wiki/C4.5_algorithm

which "learn" rules that look more like ordinary code.

There are some older algorithms like <snip> which "learn" rules that look more like ordinary code.

See also, "Rule Induction"[1].

[1]: https://en.wikipedia.org/wiki/Rule_induction

No. Macros are compile-time.

Lisp implementations typically give running programs access to the compiler, but that has nothing to do with macros, and then again, other languages do as well.

If you can compile code at runtime, then how are you not able to use macros at runtime via the same facility? Macros are just functions that operate on the data structures that define function calls prior to those calls being executed, symbols resolved , etc. So if you’re compiling new code at runtime via eval the that code will go through the same process, including macro expansion.

The point was that macros are not special, they don't give you anything more for "self-modification".

Macros are special in the sense that they are expressly intended to modify code. I take your implied point that at runtime there is not really a difference between a function designed to rewrite code data structures and evaluate them vs calling macroexpand and using a predefined macro. But it’s going to be more concise at least to do the latter. It’s also probably the case that when people talk about lisp macros they are including facilities that can be used anywhere but are mainly used for macros - backtick quoting, unquoting forms, etc. - as well as macro like activities like having a function designed to rewrite a code data structure at runtime prior to it going to eval (even if defined with defn or defun rather than defmacro).

> Macros are special in the sense that they are expressly intended to modify code

Macros are intended to transform code. They are designed such that in compiled code the transformation is only done at compile time.

If you are going to use EVAL at runtime, then macros don't buy us anything new.

calling (eval (generate-some-code)) does already execute newly generated code at runtime.

Ruby code can be generated and evaluated in runtime.

Reply on news.ycombinator.com